Sanaf, the AI employee

An AI employee that can use the web

Some of the tools your business depends on have no API worth the name: the supplier portal, the council site, the old system nobody will replace. Sanaf can sign in and work in the page itself, which is slower and more fragile than an API, and is treated accordingly.

What it does

Concretely

  • Signs in and works in the page when a tool has no API.
  • Says in the work log when it was working in the browser rather than an API.
  • Asks more often in the browser, because a page is easier to misread than a response.
  • Credentials for a site are scoped to the job that needs them.
  • Anything it cannot undo still goes through the same approval rules.

Why this exists

Integration catalogues are written for the software people wish businesses ran on. Real businesses have a supplier portal from 2011, a council website with a session timeout, and one system that only the office manager understands. None of those have an API, and none of them are going away.

What it actually does

It opens the site, signs in with credentials you gave it for that job, navigates to the thing, reads it or fills it in, and comes back with what it found or what it did. From your side it is the same as any other request; the difference is only in how it got there.

Why it is treated more carefully

A page is easier to misread than a structured response. A layout changes, a button moves, a confirmation dialog appears that was not there last week. So work in the browser asks for confirmation more often, and it always says in the work log that it was in the browser, so you can weigh what it tells you accordingly.

Credentials

A site login is a connection like any other: scoped to the job that needs it, encrypted at rest, and revocable on its own without touching anything else. If the site offers a limited account, use one. The rule is the same as it would be for a new starter, for the same reasons.

Questions

Frequently asked

Can it use any website?
Most of them, but not reliably all. Sites behind hard bot protection, or ones that require a code sent to a device, are where it stops and asks you. It tells you that rather than quietly failing.
Is working in a browser as reliable as an API?
No, and we would rather say so. A page can change under it in ways a documented API cannot. That is exactly why it asks more often when it is working this way, and why the work log always says which it used.
Where are the site logins kept?
Encrypted, scoped to the job that needs them, and revocable one at a time from the console. Revoking a site login never affects any other connection.